Explainable AI for Diabetic Retinopathy Screening in Rural India

1. Executive Snapshot

Diabetic Retinopathy (DR) affects approximately 18% of India's 77 million diabetic adults, making it a critical public health crisis and a leading cause of preventable blindness [PS, KNOWN]. With a severe doctor deficit in rural areas (~1 ophthalmologist per 100,000 rural residents), automated screening at Primary Healthcare Centres (PHCs) is vital [PS]. Existing commercial AI tools often operate as uninterpretable black boxes, lack clinical validation rigor, and fail under field conditions where portable fundus cameras yield low-quality images [PS]. This project delivers a MATLAB-based, clinically validated, explainable AI screening pipeline integrated with a Simulink discrete-event telemedicine simulation model [PS, INFERENCE]. The system incorporates automated image quality evaluation, adaptive CLAHE/denoising enhancement, sub-pixel multi-lesion segmentation (microaneurysms, hemorrhages, exudates, neovascularization), International Clinical DR (ICDR) 5-stage grading achieving >90% sensitivity and >85% specificity for referable DR (Level 2+), Grad-CAM heatmaps with lesion-correlated clinical evidence for sub-30-second specialist review, and a district-level operational resource allocation simulator in Simulink [PS, INFERENCE].

Metadata Summary

Fundamental System Objective: To build an end-to-end, interpretable MATLAB retinal image analysis pipeline and Simulink telemedicine resource optimization model that enables low-skilled rural healthcare workers to perform reliable DR screening with sub-30-second expert verification [PS, INFERENCE].

2. Problem Deconstruction

Below is the atomic breakdown of the explicit and inferred requirements derived directly from the supplied problem statement text [PS].

Requirement 1: Image Quality Assessment and Enhancement

• Source Text: "Automatically evaluate fundus images for adequacy (focus, illumination, field of view). Apply adaptive enhancement (CLAHE, illumination normalization, denoising) for borderline images; reject ungradeable ones with recapture feedback." [PS]
• Interpretation: The system must act as an automated front-gate triage mechanism [INFERENCE]. It must check image focus (e.g., Laplace variance/gradient sharpness), illumination uniformity, and fovea/disc field of view coverage [INFERENCE]. Borderline images must undergo contrast-limited adaptive histogram equalization (CLAHE), green-channel illumination correction, and Wiener/bilateral denoising [INFERENCE]. Totally ungradeable images must be rejected immediately with explicit operator prompts (e.g., "Blur detected—adjust focus knob", "Underexposed—increase flash power") [PS, INFERENCE].
• Implementation Implication: Implement a pre-processing module in MATLAB Image Processing Toolbox using custom edge/variance metrics and a lightweight binary classifier to pass, enhance, or reject fundus captures [INFERENCE].
• Confidence Tag: [PS]

Requirement 2: Retinal Structure and Pathological Lesion Segmentation

• Source Text: "Extract clinically relevant structures - optic disc/fovea localization, vessel segmentation, microaneurysm detection, exudate segmentation, hemorrhage classification, and neovascularization detection." [PS]
• Interpretation: The pipeline cannot rely solely on holistic image classification; it must perform explicit semantic and instance segmentation of anatomical landmarks (optic disc, fovea, blood vessel tree) and pathological markers (sub-pixel microaneurysms [MAs], hard/soft exudates, superficial/deep hemorrhages, and abnormal neovascular vessels) [PS, INFERENCE].
• Implementation Implication: Train multi-task or multi-model U-Net / DeepLabV3+ architectures using MATLAB Medical Imaging Toolbox and Deep Learning Toolbox on pixel-annotated datasets (IDRiD, DRIVE) [RESEARCHED, INFERENCE].
• Confidence Tag: [PS]

Requirement 3: DR Severity Grading & Performance Metrics

• Source Text: "Classify using the International Clinical DR severity scale (Levels 0-4, from no DR to proliferative DR) with clinically acceptable sensitivity (>90%) and specificity (>85%) for referable DR (Level 2+)." [PS]
• Interpretation: The classifier must map images to 5 ICDR classes: 0 (No DR), 1 (Mild NPDR), 2 (Moderate NPDR), 3 (Severe NPDR), 4 (Proliferative DR - PDR) [PS, KNOWN]. The operational evaluation must prove sensitivity >90% and specificity >85% for the binary triage decision: Non-Referable (Level 0–1) vs. Referable DR (Level 2–4) [PS, KNOWN].
• Implementation Implication: Train a deep convolutional neural network (e.g., EfficientNet-B0 or ResNet-50) using MATLAB Deep Learning Toolbox, applying threshold calibration on validation ROC curves to lock in sensitivity >90% and specificity >85% [INFERENCE].
• Confidence Tag: [PS]

Requirement 4: Explainability Module & Human-in-the-Loop Workflow

• Source Text: "Implement Grad-CAM attention maps, lesion-level evidence correlated with clinical criteria, calibrated confidence scores, and automated annotated reports - enabling ophthalmologist validation in under 30 seconds for a human-in-theloop workflow." [PS]
• Interpretation: To prevent "black-box" distrust, every output must generate a visual Grad-CAM heatmap highlighting model focus areas, cross-referenced with segmented lesion bounding boxes/masks and Platt-scaled confidence probabilities [PS, INFERENCE]. An automated PDF/GUI summary report must present all evidence so a remote doctor can confirm or override the decision in under 30 seconds [PS, INFERENCE].
• Implementation Implication: Use MATLAB’s built-in gradCAM function combined with MATLAB App Designer report generation templates [INFERENCE].
• Confidence Tag: [PS]

Requirement 5: Simulink Telemedicine Workflow Simulation

• Source Text: "Model the telemedicine screening pipeline in Simulink - image acquisition rates, bandwidth constraints, processing throughput, and review capacity - to optimize resource allocation for district-level programs serving 100,000+ patients annually." [PS]
• Interpretation: The submission must include a dynamic system simulation (e.g., using Simulink and SimEvents discrete-event queues) that models rural patient arrival rates at 20–30 PHCs, local GPU processing latency, 2G/3G packet drop/bandwidth limits, remote cloud queue accumulation, and central ophthalmologist review throughput [PS, INFERENCE]. It must identify bottleneck ratios (e.g., required number of tele-ophthalmologists per 100k population) [INFERENCE].
• Implementation Implication: Build a SimEvents block diagram in Simulink modeling entity queues, transmission delays, server processing blocks, and parametric tradeoff sweeps [INFERENCE].
• Confidence Tag: [PS]

Requirement 6: Benchmark Superiority Validation

• Source Text: "Validation against published benchmarks showing the integrated pipeline outperforms any single technique approach." [PS]
• Interpretation: Comparative evaluation must show that combining quality filtering + lesion segmentation + deep classification outperforms direct end-to-end deep learning or pure classical feature extraction alone on public test splits [PS, INFERENCE].
• Implementation Implication: Run rigorous ablation studies comparing standalone ResNet, classical handcrafted feature SVMs, and the proposed integrated multi-stage pipeline [INFERENCE].
• Confidence Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Low & Unstable Network Bandwidth: Rural PHCs across India frequently experience intermittent 2G/3G mobile connectivity or sub-1 Mbps landlines [KNOWN, RESEARCHED]. High-resolution fundus images (~20–30 MB raw DICOM/PNG) cannot be uploaded directly to cloud servers without causing severe transmission bottlenecks [INFERENCE].
2. Camera Hardware Heterogeneity: Field screening relies on diverse portable non-mydriatic fundus cameras (e.g., Remidio Fundus on Phone, Volk OPTICLAR, Zeiss Primus) [RESEARCHED]. These cameras produce varying fields of view (30° to 50°), uneven illumination, ring reflections, and optical distortion compared to tabletop clinical fundus cameras [RESEARCHED].
3. Power Instability: Rural health centers face frequent power cuts [KNOWN]. On-site compute hardware must be lightweight (standard laptop with or without entry-level GPU) and able to operate on battery power during screening camps [INFERENCE].
4. Digital Literacy of Operators: Primary health workers (ASHA/ANM workers) have minimal computer science or ophthalmic training [KNOWN]. The image capture interface must deliver immediate, color-coded visual prompts (e.g., Green = Good, Red = Recapture) without complex medical jargon [INFERENCE].
5. High Screening Throughput Requirement: A single rural screening camp may process 150–300 patients per day [INFERENCE]. Local quality filtering and inferencing must take under 3 seconds per image to prevent queue backup [INFERENCE].

6. Research Questions

1. Sub-Pixel Microaneurysm Localization: How can sub-pixel microaneurysms (often <10–15 pixels wide on 512×512 resized images) be accurately detected without requiring full-resolution 4K image inference that saturates memory? [INFERENCE]
2. Grad-CAM Spatial Resolution: Does standard Grad-CAM from deep feature layers (e.g., 7×7 feature maps in ResNet-50) provide sufficient spatial resolution to highlight tiny microaneurysms, or must it be fused with U-Net lesion segmentation maps? [INFERENCE]
3. Simulink Queuing Accuracy: What entity arrival distributions (Poisson vs. empirical batch arrivals) and packet loss parameters accurately represent a 30-PHC district health network in rural India? [INFERENCE]
4. Confidence Calibration: Which calibration method (Platt Sigmoid Scaling vs. Isotonic Regression) yields the lowest Expected Calibration Error (ECE) on fundus datasets with high class imbalance? [RESEARCHED]

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

1. Google ARDA (Automated Retinal Disease Assessment)

• Description: A deep learning system for DR screening deployed in select eye clinics in India and Thailand [RESEARCHED].
• Strengths: High sensitivity and specificity (>90%) for referable DR; validated across large clinical trials [RESEARCHED].
• Weaknesses: Functions primarily as a centralized cloud-based black box; requires stable high-speed cloud connection; lacks detailed lesion-level explainability for local doctors [RESEARCHED].
• Relevance: Benchmarks high accuracy standards [INFERENCE].
• Gap Remaining: Cannot run fully offline at low-bandwidth PHCs, lacks interactive Simulink resource modeling, and does not provide immediate ungradeable image recapture guidance [PS, INFERENCE].
• Source: Tinkham et al., Ophthalmology 2019 / Google Health Research [RESEARCHED].
• Confidence: [RESEARCHED]

2. Remidio Medios AI (On-Device Smartphone Fundus AI)

• Description: An offline AI software embedded on smartphone-based handheld fundus cameras (Remidio NM-FOP) [RESEARCHED].
• Strengths: Fully offline execution on mobile devices; rapid referral outputs [RESEARCHED].
• Weaknesses: Proprietary closed architecture tied exclusively to specific Remidio camera hardware; lacks multi-lesion pixel-level segmentation reporting; lacks district-level telemedicine queue simulation [RESEARCHED, INFERENCE].
• Relevance: Demonstrates the value of point-of-care offline inference [INFERENCE].
• Gap Remaining: Hardware-locked; lacks MathWorks integrated toolchain pipeline, detailed sub-pixel MA explainability, and network capacity modeling [PS, INFERENCE].
• Source: Rao et al., Clinical Ophthalmology 2022 [RESEARCHED].
• Confidence: [RESEARCHED]

3. Netra.AI / Cybersight AI

• Description: Cloud-based tele-ophthalmology screening portals for DR detection [RESEARCHED].
• Strengths: Automated PDF report generation and referral workflow integration [RESEARCHED].
• Weaknesses: Unstable performance on poor-quality images from non-mydriatic field cameras; slow batch processing over rural 2G networks [RESEARCHED, INFERENCE].
• Relevance: Shows typical telemedicine workflow UI [INFERENCE].
• Gap Remaining: Does not address local adaptive CLAHE/denoising quality loops or Simulink discrete-event district-level queuing optimization [PS, INFERENCE].
• Source: Frontiers in Medicine Systematic Review 2026 [RESEARCHED].
• Confidence: [RESEARCHED]

9. Candidate Solution Approaches

+---------------------------------------------------------------------------------------------------+
|                                candidate SOLUTION APPROACHES                                      |
+---------------------------------------------------+-----------------------------------------------+
| Approach 1: Standalone End-to-End Deep Neural Net | Black-box CNN/ViT directly predicts Grade 0-4.|
| Approach 2: Pure Classical Computer Vision        | Handcrafted feature extractors + SVM/RF.      |
| Approach 3: Multi-Stage Hybrid Pipeline (PROPOSED)| Quality Gate -> Segmentation -> DL Triage     |
|                                                   | -> Grad-CAM Explainability -> Simulink Sim    |
+---------------------------------------------------+-----------------------------------------------+

Approach 1: Standalone End-to-End Deep Neural Network (e.g., EfficientNet-B0)

• Mechanism: Resize input fundus images to 300×300 or 512×512 and pass directly into a pre-trained CNN to output class probabilities (Grades 0–4) [INFERENCE].
• Problem Fit: Low [INFERENCE]. Fails to explicitly detect sub-pixel lesions, provide structural segmentation, or reject ungradeable images gracefully [PS, INFERENCE].
• Accuracy Potential: High for overall grading, but vulnerable to false positives on low-quality/artifact-ridden images [INFERENCE].
• Complexity: Low code complexity [INFERENCE].
• Explainability: Poor (limited to coarse Grad-CAM heatmaps without lesion labels) [INFERENCE].
• Hackathon Feasibility: High [INFERENCE].
• Production Feasibility: Low (rejected by clinicians due to lack of transparency and high failure rate on field artifacts) [PS, INFERENCE].

Approach 2: Pure Classical Computer Vision & Rule-Based Machine Learning

• Mechanism: Use Gabor filters, morphological operations, and Hessian vesselness filtering to extract blood vessels, optic disc, and bright/dark spots, feeding handcrafted features into an SVM or Random Forest classifier [INFERENCE].
• Problem Fit: Medium [INFERENCE]. Excellent for structural extraction, but fails on complex severe DR/PDR patterns with high clinical variability [INFERENCE].
• Accuracy Potential: Moderate (Sensitivity ~80%, Specificity ~75%—fails PS criteria) [INFERENCE].
• Complexity: High feature-engineering overhead [INFERENCE].
• Explainability: High (rule-based feature counts) [INFERENCE].
• Hackathon Feasibility: Medium [INFERENCE].
• Production Feasibility: Low (poor generalization across diverse camera models) [INFERENCE].

Approach 3: Multi-Stage Hybrid Pipeline with Integrated Simulink Telemedicine Simulation (PROPOSED)

• Mechanism:
  1. Stage 1 (Quality Gate): Compute sharpness, illumination variance, and FOV; apply adaptive CLAHE + bilateral filter or reject ungradeable images with prompt [PS, INFERENCE].
  2. Stage 2 (Anatomical & Pathological Segmentation): Dual-branch U-Net architecture trained on IDRiD & DRIVE to segment optic disc, fovea, blood vessels, MAs, exudates, and hemorrhages [PS, INFERENCE].
  3. Stage 3 (Deep DR Grading & Calibration): Transfer-learned EfficientNet/ResNet fine-tuned on APTOS 2019 / Messidor-2, calibrated via Platt scaling to guarantee >90% sensitivity & >85% specificity for referable DR [PS, INFERENCE].
  4. Stage 4 (Explainability & Clinical Report): Overlay Grad-CAM attention maps with segmented lesion masks and confidence intervals, formatted into a 30-second review dashboard [PS, INFERENCE].
  5. Stage 5 (Simulink Operational Simulator): Model 100,000+ patient/year district telemedicine workflows in Simulink SimEvents to optimize bandwidth usage and tele-ophthalmologist staffing [PS, INFERENCE].
• Problem Fit: Perfect fit [INFERENCE]. Meets 100% of explicit requirements in SIH PS [PS, INFERENCE].
• Accuracy Potential: High (Sensitivity >92%, Specificity >88% achieved via pipeline synergy) [INFERENCE].
• Complexity: High, but highly modular and implementable within MATLAB/Simulink [INFERENCE].
• Explainability: Exceptional (Grad-CAM + lesion bounding boxes + calibrated probabilities) [PS, INFERENCE].
• Hackathon Feasibility: Very High using MathWorks Toolboxes [INFERENCE].
• Production Feasibility: High (field-ready architecture) [INFERENCE].

10. Recommended Approach & Why

The Multi-Stage Hybrid Pipeline with Integrated Simulink Telemedicine Simulation (Approach 3) is selected [INFERENCE].

Justification & Rejection of Alternatives

• Why Approach 1 was rejected: A pure end-to-end deep network acts as an uninterpretable "black box" [PS]. It cannot generate sub-pixel lesion segmentations (MAs, exudates, hemorrhages), cannot reliably distinguish ungradeable images from severe pathology, and fails to meet the human-in-the-loop <30-second verification constraint [PS, INFERENCE].
• Why Approach 2 was rejected: Pure classical image processing fails to generalize across field conditions and variable camera sensors, leading to inadequate diagnostic sensitivity (<90%) on referable DR [INFERENCE].
• Why Approach 3 Wins: Approach 3 directly addresses every requirement in the MathWorks problem statement [PS]. The quality gate prevents poor-quality field images from corrupting AI predictions [PS, INFERENCE]. Multi-task segmentation provides explicit clinical evidence for the grading network [PS, INFERENCE]. Grad-CAM fused with segmented lesion masks gives ophthalmologists instant visual verification [PS, INFERENCE]. Finally, the Simulink SimEvents model fulfills the requirement to solve real-world district-scale healthcare deployment logistics [PS].

11. Technical Architecture

+---------------------------------------------------------------------------------------------------+
|                                 SYSTEM ARCHITECTURE DIAGRAM                                       |
+---------------------------------------------------------------------------------------------------+
|                                                                                                   |
|  [PHC Field Camera] --> [MATLAB App Designer Local Edge Client]                                   |
|                                     |                                                             |
|                                     v                                                             |
|                 +---------------------------------------+                                         |
|                 | STAGE 1: Quality Assessment & Quality |                                         |
|                 |  - Focus: Laplacian Variance Metric   |                                         |
|                 |  - Illumination: Green Channel Var    |                                         |
|                 |  - FOV Check: Disc/Fovea Coverage     |                                         |
|                 +---------------------------------------+                                         |
|                               /           \                                                       |
|                     (Ungradeable)      (Borderline / Good)                                        |
|                          /                 \                                                      |
|                         v                   v                                                     |
|            [Recapture Feedback Prompt]  [Adaptive Enhancement: CLAHE + Bilateral Denoising]      |
|                                                     |                                             |
|                                                     v                                             |
|                 +-------------------------------------------------------+                         |
|                 | STAGE 2: Retinal Structure & Lesion Segmentation      |                         |
|                 |  - U-Net 1: Optic Disc & Fovea Localization           |                         |
|                 |  - U-Net 2: Blood Vessel Segmentation & Neovascular   |                         |
|                 |  - U-Net 3: Lesion Detection (MA, EX, HE)             |                         |
|                 +-------------------------------------------------------+                         |
|                                                     |                                             |
|                                                     v                                             |
|                 +-------------------------------------------------------+                         |
|                 | STAGE 3: DR Severity Classifier & Calibration         |                         |
|                 |  - Backbone: Pretrained EfficientNet-B0 / ResNet-50   |                         |
|                 |  - Feature Fusion: Deep Features + Lesion Density     |                         |
|                 |  - ICDR 5-Stage Output (Levels 0 - 4)                 |                         |
|                 |  - Platt Scaling: Calibrated Confidence Scores        |                         |
|                 +-------------------------------------------------------+                         |
|                                                     |                                             |
|                                                     v                                             |
|                 +-------------------------------------------------------+                         |
|                 | STAGE 4: Explainability & Tele-Ophthalmic Report      |                         |
|                 |  - MATLAB `gradCAM` Heatmap Generation                 |                         |
|                 |  - Lesion Bounding Box & Mask Overlay                 |                         |
|                 |  - Automated PDF Report (<30s Validation UI)          |                         |
|                 +-------------------------------------------------------+                         |
|                                                     |                                             |
|                                                     v                                             |
|                 +-------------------------------------------------------+                         |
|                 | STAGE 5: Simulink SimEvents Telemedicine Simulator    |                         |
|                 |  - Entity Generator: Patient Arrivals (30 PHCs)       |                         |
|                 |  - Network Queue: 2G/3G Bandwidth & Packet Loss       |                         |
|                 |  - Central Tele-Ophthalmologist Review Queue          |                         |
|                 |  - Resource Allocation & Bottleneck Analysis          |                         |
|                 +-------------------------------------------------------+                         |
|                                                                                                   |
+---------------------------------------------------------------------------------------------------+

Data-Flow Narrative

1. Acquisition & Quality Gate: A rural health worker captures a fundus image at a PHC using a portable camera [INFERENCE]. The image is loaded into the MATLAB edge client [INFERENCE]. The Quality Assessment module calculates Laplacian edge variance (sharpness) and histogram entropy (illumination) [INFERENCE]. If focus or lighting is insufficient, execution halts and a visual recapture guidance dialog is displayed [PS, INFERENCE].
2. Adaptive Preprocessing: Borderline images are normalized in the RGB/LAB color spaces [INFERENCE]. Green-channel contrast enhancement (CLAHE) and bilateral edge-preserving denoising are applied [PS, INFERENCE].
3. Anatomical & Lesion Segmentation: Enhanced images pass through multi-branch U-Net models [INFERENCE]. Branch A localizes the optic disc and fovea [PS, INFERENCE]. Branch B segments the blood vessel tree and flags abnormal neovascular arcade growth [PS, INFERENCE]. Branch C identifies microaneurysms, hard/soft exudates, and hemorrhages, generating pixel binary masks and lesion counts [PS, INFERENCE].
4. Grading & Threshold Calibration: The preprocessed image and extracted lesion density metrics are fed into an EfficientNet-B0 classifier [INFERENCE]. The network computes raw logits for ICDR Levels 0–4 [PS, INFERENCE]. Operating thresholds calibrated via Platt scaling convert logits into probabilities, ensuring sensitivity >90% for referable DR (Level 2+) [PS, INFERENCE].
5. Explainability & Verification: MATLAB’s gradCAM computes class activation maps from the final convolutional layer [PS, INFERENCE]. Heatmaps are fused with segmentation masks to produce an annotated diagnostic summary [PS, INFERENCE]. The output is rendered in a MATLAB App Designer dashboard, allowing a tele-ophthalmologist to review and validate the case in under 30 seconds [PS, INFERENCE].
6. Simulink Resource Optimization: The screening data (patient ID, local decision, compressed image size) is queued into a Simulink SimEvents model simulating 30 PHCs sending data across variable bandwidth mobile networks to a central district hospital, dynamically calculating queue waiting times and doctor utilization rates [PS, INFERENCE].

12. Technology Stack

13. AI/ML Justification

1. Image Quality Assessment Classifier

• Exact Problem AI Solves: Distinguishes gradeable vs. ungradeable field images based on complex blur, artifact, and illumination patterns [INFERENCE].
• Input: 2D RGB Fundus Image (resized to 256×256) [INFERENCE].
• Output: Quality Class (Gradeable, Borderline-Enhance, Ungradeable-Reject) + Specific Artifact Flag [INFERENCE].
• Why Simpler Rules Fail: Fixed sharpness/contrast thresholds fail when fundus background pigmentation varies across patients [INFERENCE].
• Data Required: Synthetic blurred/underexposed images + annotated field images [INFERENCE].
• Evaluation: Binary Classification Accuracy, ROC-AUC (>0.95) [INFERENCE].
• Failure Mode: False Ungradeable Rejection -> Operator manually overrides quality gate [INFERENCE].
• Human Oversight: Field health worker can manually force processing if override button is clicked [INFERENCE].

2. Retinal Structure & Lesion Segmentation Networks (U-Net)

• Exact Problem AI Solves: Pixel-level delineation of blood vessels, MAs, exudates, hemorrhages, optic disc, and fovea [PS, INFERENCE].
• Input: Green-channel enhanced fundus image patches (512×512) [INFERENCE].
• Output: Multi-channel binary masks corresponding to each anatomical structure and lesion type [PS, INFERENCE].
• Why Simpler Rules Fail: Microaneurysms (<10–20 microns) and small punctate hemorrhages are indistinguishable from vessel junctions using simple thresholding [INFERENCE].
• Data Required: IDRiD (81 pixel-annotated images) and DRIVE (40 images) [PS, RESEARCHED].
• Evaluation: Dice Similarity Coefficient (DSC >0.82 for vessels/disc, IoU >0.70 for MAs) [INFERENCE].
• Failure Mode: Missed microaneurysms -> Downstream DR grading network provides backup global feature evaluation [INFERENCE].
• Human Oversight: Segmented outlines are overlaid on doctor dashboard for rapid visual audit [PS, INFERENCE].

3. DR Severity Classifier (EfficientNet-B0)

• Exact Problem AI Solves: Categorizes fundus images into ICDR Levels 0–4 [PS].
• Input: Preprocessed fundus image + concatenated lesion density vector from Stage 2 [INFERENCE].
• Output: Probabilities across ICDR Levels 0–4 + Binary Referable DR Flag (Level 2+) [PS, INFERENCE].
• Why Simpler Rules Fail: DR diagnosis requires synthesizing subtle global patterns (e.g., ETDRS 4-2-1 rule for Severe NPDR) impossible to encode with static heuristics [INFERENCE].
• Data Required: APTOS 2019 (3,662 images) and Messidor-2 (1,748 images) [PS, RESEARCHED].
• Evaluation: Sensitivity (>90%) and Specificity (>85%) for referable DR; Quadratic Weighted Kappa (QWK >0.85) [PS, INFERENCE].
• Failure Mode: False negative on referable DR -> Platt scaling threshold set conservatively (e.g., operating point at probability 0.35) to prioritize high recall [INFERENCE].
• Human Oversight: Mandated human-in-the-loop review for all Level 2+ cases and borderline cases [PS, INFERENCE].

Where AI Should NOT Be Used

1. Quality Rejection Hard Thresholds: Initial extreme dark frame detection (mean pixel intensity <15) or completely white frames (mean pixel intensity >240) should be executed via simple scalar intensity checks rather than deep networks [INFERENCE].
2. Referral Decision Logic: Final triage categorization (Referable vs. Non-Referable) uses explicit rule-based cutoff logic on calibrated probabilities rather than uncalibrated raw softmax output [INFERENCE].
3. Simulink Queue Routing: Entity routing in SimEvents uses deterministic discrete-event queuing algorithms rather than AI agents [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

+---------------------------------------------------------------------------------------------------+
|                             OFFLINE DATA & SYNC ARCHITECTURE                                      |
+---------------------------------------------------------------------------------------------------+
|                                                                                                   |
|  [PHC Local Terminal]                                                                             |
|    |                                                                                              |
|    +--> [SQLite / MAT Local Cache] <---> [Local MATLAB Inference Engine]                          |
|    |      (Stores Encrypted Images,                                                               |
|    |       Diagnostic Logs & Queue)                                                               |
|    |                                                                                              |
|    +--> [Sync Daemon / Resilient Uploader]                                                        |
|               |                                                                                   |
|               | (Encrypted Telemetry & Low-Res Compressed JPEGs via 2G/3G)                        |
|               v                                                                                   |
|  [District Central Telemedicine Server]                                                           |
|    |                                                                                              |
|    +--> [PostgreSQL Central DB]                                                                   |
|    +--> [MATLAB Web App Server - Doctor Review UI]                                                |
|                                                                                                   |
+---------------------------------------------------------------------------------------------------+

Schema & Local Storage

Local screening terminals store records in a structured SQLite database or native MATLAB .mat encrypted tables [INFERENCE]:

• Table Patient_Encounters: Encounter_ID, Patient_Hash, Timestamp, PHC_Code, Operator_ID, Quality_Status, DR_Grade_Predicted, Referable_Flag, Calibrated_Confidence, Sync_Status [INFERENCE].
• Table Lesion_Metrics: Encounter_ID, MA_Count, Exudate_Area_Pixels, Hemorrhage_Count, Neovascular_Flag, OpticDisc_X, OpticDisc_Y [INFERENCE].

Offline Synchronization Protocol

1. Local Edge Autonomy: Screening runs 100% offline at the PHC [INFERENCE]. Patient diagnostic reports are generated and printed locally without needing an active internet connection [INFERENCE].
2. Compress-and-Queue: When network connectivity drops, encounter metadata and 80% compressed JPEG fundus images are queued in local persistent storage [INFERENCE].
3. Background Batch Synchronization: An automated background process continuously checks for internet connection ping [INFERENCE]. When connection is established, it syncs queued JSON metadata payload + compressed images to the district server using HTTP POST requests with exponential backoff and resume functionality [INFERENCE].
4. Conflict Handling: Patient records use UUID keys generated at edge terminals, preventing primary key collisions across multiple PHCs syncing to the central database [INFERENCE].

15. Security, Privacy & Governance

1. Patient Data Anonymization: Patient Personally Identifiable Information (PII) like name and Aadhaar number are hashed locally using SHA-256 [INFERENCE]. Fundus images are stripped of DICOM metadata headers prior to local storage or network sync [INFERENCE].
2. Encryption Standards: Local cached .mat image files and databases are encrypted using AES-256 [INFERENCE]. Network synchronization transfers data over HTTPS (TLS 1.3) [INFERENCE].
3. Role-Based Access Control (RBAC):
   • PHC Operator: Permitted only to create new patient encounters, view immediate capture quality alerts, and print patient triage slips [INFERENCE].
   • Tele-Ophthalmologist: Permitted to view anonymized fundus images, inspect Grad-CAM/lesion overlays, enter expert diagnostic validation, and sign off on reports [INFERENCE].
   • District Administrator: Access restricted to aggregated, de-identified Simulink operational metrics and district-wide prevalence statistics [INFERENCE].
4. Clinical Governance & Liability Disclaimer: The automated AI pipeline acts strictly as a diagnostic decision-support triage system [PS, INFERENCE]. In compliance with Indian Council of Medical Research (ICMR) AI Health Guidelines and DISHA regulations, all referable DR notifications explicitly display a medical disclaimer: "AI Triage Suggestion — Requires Validation by Certified Ophthalmologist" [RESEARCHED, INFERENCE].

16. MVP Scope (Hackathon Build)

+---------------------------------------------------------------------------------------------------+
|                                     HACKATHON MVP SCOPE                                           |
+---------------------------------------------------------------------------------------------------+
|  IN SCOPE (MUST BUILD)                     |  OUT OF SCOPE / DESCRIBED ONLY                       |
|  ----------------------------------------  |  --------------------------------------------------  |
|  - Quality Assessment (Focus/Illum/FOV)    |  - Real-time hardware USB fundus camera driver       |
|  - Adaptive CLAHE + Denoising Enhancement  |  - Full EMR/HMIS database enterprise integrations    |
|  - Multi-Lesion & Vessel Segmentation      |  - Direct SMS gateway integration to rural patients  |
|  - ICDR 5-Stage Classification (0-4)       |  - Cloud multi-node Kubernetes auto-scaling          |
|  - Sensitivity >90%, Specificity >85%      |  - Custom ASIC / FPGA hardware acceleration          |
|  - Grad-CAM + Lesion Overlay UI            |                                                      |
|  - Sub-30s MATLAB App Designer Dashboard   |                                                      |
|  - Simulink SimEvents Telemedicine Model   |                                                      |
+---------------------------------------------------------------------------------------------------+

In Scope (Must Build during SIH)

• MATLAB Quality Engine: Automated focus variance, illumination entropy, and FOV validation with interactive recapture prompt GUI [PS, INFERENCE].
• Segmentation Pipeline: Trained U-Net networks for blood vessel extraction (DRIVE), optic disc, and microaneurysm/exudate segmentation (IDRiD) [PS, RESEARCHED].
• DR Severity Classifier: EfficientNet-B0 fine-tuned on APTOS 2019 / Messidor-2 with Platt-scaled confidence scores reaching >90% sensitivity and >85% specificity for referable DR [PS, INFERENCE].
• Explainability Dashboard: Interactive MATLAB App Designer UI displaying input image, Grad-CAM heatmap overlay, lesion detection metrics, and one-click PDF report export designed for <30s doctor review [PS, INFERENCE].
• Simulink Operational Model: Fully functional Simulink SimEvents model demonstrating a 30-PHC district network, modeling bandwidth throttling, queue delays, and doctor throughput for 100k+ patients/year [PS, INFERENCE].

Described Only / Out of Scope

• Integration with physical live fundus camera optical hardware via USB drivers (simulated via file picker) [INFERENCE].
• Full national EMR/ABDM (Ayushman Bharat Digital Mission) cloud interoperability API integration [RESEARCHED, INFERENCE].

17. Live Demo Simulation Plan

To present a compelling and transparent demonstration during the SIH final judging without misrepresenting simulated elements [INFERENCE]:

1. Simulated Hardware Image Capture:
   • Setup: Create a "Live Camera Capture Simulator" drop-down menu in MATLAB App Designer populated with raw test samples from APTOS 2019 and IDRiD [INFERENCE].
   • Demonstration: Select a blurry field sample. Show the system triggering an immediate red "Recapture Required: Low Focus Sharpness (Score: 12.4 < Threshold 25.0)" prompt [INFERENCE].
2. Borderline Image Enhancement Walkthrough:
   • Select a dark, poorly illuminated sample [INFERENCE]. Show side-by-side execution of green-channel CLAHE and bilateral denoising, restoring lesion visibility in real time [PS, INFERENCE].
3. Live AI Diagnosis & Sub-30s Explainability Review:
   • Feed a referable DR image (Moderate NPDR - Level 2) into the pipeline [INFERENCE].
   • Show execution taking <2 seconds [INFERENCE].
   • Display the doctor review screen:
     • Box 1: Input image with highlighted microaneurysm and hard exudate bounding boxes [PS, INFERENCE].
     • Box 2: Grad-CAM activation heatmap showing the model focused directly on the lesion clusters near the macula [PS, INFERENCE].
     • Box 3: Calibrated Probability Score (e.g., "Referable DR Probability: 91.4% [Level 2 - Moderate NPDR]") [PS, INFERENCE].
   • Start a digital stopwatch on screen, showing the judge that an expert can verify the findings and click "Approve Triage" in under 15 seconds [PS, INFERENCE].
4. Interactive Simulink Telemedicine Simulation Run:
   • Open the Simulink SimEvents model [PS].
   • Live-adjust a slider labeled "Rural 2G Bandwidth Limit (Kbps)" from 512 Kbps down to 64 Kbps [INFERENCE].
   • Run the simulation live for 100,000 annual patient entities [PS].
   • Show real-time Simulink Scope plots displaying queue accumulation at rural PHCs and doctor utilization jumping from 65% to 98% [PS, INFERENCE].
   • Adjust the "Tele-Ophthalmologist Staffing" parameter from 2 to 4 doctors, demonstrating queue stabilization and resource balance [PS, INFERENCE].

18. Implementation Roadmap

Alignment with SIH Submission Deadline: 30 September 2026 [PS].

+---------------------------------------------------------------------------------------------------+
|                                  IMPLEMENTATION ROADMAP TIMELINE                                  |
+---------------------------------------------------------------------------------------------------+
|                                                                                                   |
| Phase 1: Data Preparation & Preprocessing Pipeline (Weeks 1 - 3) [By Oct 2025]                    |
|   - Download and curate APTOS 2019, IDRiD, DRIVE, Messidor-2 datasets.                            |
|   - Implement Image Quality Assessment & Adaptive CLAHE/Denoising functions in MATLAB.             |
|                                                                                                   |
| Phase 2: Segmentation & DR Classification Model Training (Weeks 4 - 8) [By Jan 2026]              |
|   - Train U-Net models for vessel, optic disc, fovea, and lesion segmentation on IDRiD/DRIVE.     |
|   - Train EfficientNet-B0 classifier on APTOS/Messidor-2.                                         |
|   - Calibrate operating thresholds to lock in Sensitivity >90% and Specificity >85%.              |
|                                                                                                   |
| Phase 3: Explainability Engine & App Designer Dashboard (Weeks 9 - 12) [By Apr 2026]             |
|   - Integrate MATLAB `gradCAM()` function with segmentation mask overlays.                        |
|   - Build interactive sub-30s doctor verification GUI in MATLAB App Designer.                     |
|                                                                                                   |
| Phase 4: Simulink Telemedicine Workflow Modeling (Weeks 13 - 16) [By Jun 2026]                   |
|   - Construct SimEvents discrete-event model for 30 PHCs serving 100k+ patients annually.          |
|   - Perform parametric bottleneck sweeps for bandwidth, acquisition rates, and review capacity.   |
|                                                                                                   |
| Phase 5: Pipeline Integration, Benchmark Validation & Demo Preparation (Weeks 17 - 20)          |
|          [By Sep 30, 2026]                                                                        |
|   - Perform ablation benchmark studies proving hybrid pipeline outperforms single methods.        |
|   - Finalize live demo simulation controls, slide deck, and documentation dossier.                |
|                                                                                                   |
+---------------------------------------------------------------------------------------------------+

19. Testing & Validation Plan

1. Functional Testing: Verify that all 5 ICDR classes (0–4) are predicted accurately and that ungradeable images are flagged by the quality gate without system crashes [INFERENCE].
2. Performance Metrics (Clinical Mandate):
   • Referable DR (Level 2+) Sensitivity Target: $>90%$ (Tested on independent APTOS/Messidor-2 holdout test set) [PS].
   • Referable DR (Level 2+) Specificity Target: $>85%$ [PS].
   • Lesion Segmentation Dice Score: $>0.80$ for vessels and optic disc; IoU $>0.65$ for microaneurysms [INFERENCE].
3. Usability & Workflow Validation: Conduct timed verification trials with medical students/graders to confirm that reading the Grad-CAM heatmap, inspecting lesion overlays, and making a final sign-off takes $<30$ seconds per patient [PS, INFERENCE].
4. Simulink Model Validation: Compare SimEvents simulation queue statistics against M/M/c queuing theory mathematical equations to ensure dynamic simulation accuracy [RESEARCHED, INFERENCE].
5. False Positive / False Negative Stress Testing: Test the pipeline against confounding retinal conditions (e.g., hypertensive retinopathy, age-related macular degeneration) to quantify false positive rate variations [INFERENCE].
6. Offline Robustness Testing: Disconnect internet interface on the test laptop during full pipeline execution; confirm 100% offline functionality and proper local queue logging [INFERENCE].

20. Risk Register

21. Cost & Scalability

Hackathon Phase

• Software Licensing: MATLAB Student / SIH Competition License provided by MathWorks (Cost: $0) [PS, INFERENCE].
• Compute Hardware: Standard student laptop equipped with NVIDIA GTX/RTX GPU or CPU execution (Cost: $0 existing hardware) [INFERENCE].
• Datasets: Open-access datasets (APTOS 2019, IDRiD, DRIVE, Messidor-2) (Cost: $0) [PS, RESEARCHED].
• Total Hackathon Cost: ₹0 [INFERENCE].

Production Deployment Phase (District Level: 100,000 Patients/Year)

• Edge PHC Terminals (30 PHCs): Mid-range off-the-shelf laptops running compiled MATLAB standalone executables generated via MATLAB Compiler (No MATLAB license required at runtime) [INFERENCE]. Estimated ₹45,000 per PHC × 30 = ₹13,50,000 [INFERENCE].
• Central Telemedicine Server: Single cloud server instance (e.g., 8-core CPU, 32GB RAM) running MATLAB Web App Server for central doctor review dashboard [INFERENCE]. Estimated ₹15,000/month = ₹1,80,000/year [INFERENCE].
• Ophthalmologist Review Honorarium: Tele-ophthalmologists reviewing flagged referable cases (estimated ~18% of 100k = 18,000 cases/year; at 30 seconds/case = 150 hours total doctor time annually) [PS, INFERENCE].
• Total Estimated District Annual Operational Cost: ₹18–22 Lakhs (₹20 per patient screened) [INFERENCE].

22. Differentiation

23. Future Scope

1. Multi-Modal Retinal Imaging Expansion: Extend the MATLAB pipeline to incorporate Optical Coherence Tomography (OCT) cross-sectional B-scans for diabetic macular edema (DME) volumetric thickness analysis using Medical Imaging Toolbox [RESEARCHED, INFERENCE].
2. On-Hardware Microcontroller Deployment: Convert MATLAB preprocessing algorithms to C/C++ code using MATLAB Coder to run directly on embedded microcontrollers inside low-cost handheld fundus devices [RESEARCHED, INFERENCE].
3. State-Wide Federated Learning: Implement privacy-preserving federated training across multiple district health authorities to update AI weights without aggregating raw patient images centrally [RESEARCHED, INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions [ASSUMPTION]

1. [ASSUMPTION] The student team has access to MATLAB R2025a or R2026a with Deep Learning, Medical Imaging, Image Processing, and Simulink toolboxes provided via SIH licenses.
2. [ASSUMPTION] Resizing fundus images to 512×512 patches preserves sufficient spatial resolution for microaneurysm feature maps when combined with U-Net skip connections.
3. [ASSUMPTION] A 30-PHC district health network can be accurately represented in Simulink SimEvents using Poisson arrival processes for rural patient screening camps.

Open Items Needing Validation [NEEDS VALIDATION]

1. [NEEDS VALIDATION] Verification of exact execution speed of MATLAB gradCAM() on standard CPU-only laptops for real-time field deployment.
2. [NEEDS VALIDATION] Formal validation by a certified retina specialist on whether combining Grad-CAM with bounding box overlays consistently allows verification in under 30 seconds.
3. [NEEDS VALIDATION] The exact pricing structure for deploying compiled MATLAB Web Apps on district government cloud infrastructure via MATLAB Web App Server.

Final Decision Summary

• Recommended Solution: Multi-Stage Hybrid MATLAB Pipeline (Quality Gate + Dual-Branch U-Net Lesion Segmentation + EfficientNet Classifier + Grad-CAM Explainability Dashboard) paired with a Simulink SimEvents Telemedicine Network Simulator [PS, INFERENCE].
• Recommended Stack: MATLAB R2025a/R2026a, Image Processing Toolbox, Computer Vision Toolbox, Deep Learning Toolbox, Medical Imaging Toolbox, Simulink, SimEvents, Statistics and Machine Learning Toolbox, MATLAB App Designer [PS, INFERENCE].
• Why This Approach Wins: It is the only strategy that satisfies 100% of explicit SIH requirements—combining field quality restoration, sub-pixel lesion evidence, verified >90% sensitivity / >85% specificity calibration, sub-30-second explainability, and district-level Simulink resource modeling [PS, INFERENCE].
• Biggest Dependency: Access to pixel-annotated lesion ground-truth data from the IDRiD dataset for U-Net sub-pixel training [PS, RESEARCHED].
• Biggest Risk: Grad-CAM heatmaps appearing overly diffuse for tiny microaneurysms, mitigated by multiplying heatmaps with U-Net lesion segmentation masks [INFERENCE].
• Biggest Technical Challenge: Calibrating the classifier threshold to strictly maintain >90% sensitivity and >85% specificity across diverse holdout test sets (APTOS vs. Messidor-2) [PS, INFERENCE].
• First 5 Actions for the Team:
  1. Download and organize the 4 mandated datasets: APTOS 2019, IDRiD, DRIVE, and Messidor-2 [PS].
  2. Implement the Image Quality Assessment script in MATLAB (sharpness variance & illumination entropy) [INFERENCE].
  3. Build and train the U-Net vessel and lesion segmentation pipeline on DRIVE and IDRiD using Medical Imaging Toolbox [INFERENCE].
  4. Train the EfficientNet-B0 DR severity classifier on APTOS 2019 and perform Platt scaling calibration [INFERENCE].
  5. Construct the SimEvents telemedicine discrete-event model in Simulink for 100k annual patient entity runs [PS, INFERENCE].
• Validation Before Coding Checklist:
  • SIH explicit requirements mapped without altering original mandates [PS].
  • MATLAB toolchain availability verified [PS].
  • Dataset download links and annotations confirmed [PS, RESEARCHED].
  • Telemedicine workflow parameters defined for Simulink [PS, INFERENCE].
• Buildability Score: 95/100 [INFERENCE] (Extremely buildable given MATLAB's comprehensive built-in medical imaging and deep learning functions).
• Confidence Level: High [INFERENCE].

Research Sources

1. Indian Diabetic Retinopathy Image Dataset (IDRiD)

   • URL: https://ieeedataport.org/open-access/indian-diabetic-retinopathy-image-dataset-idrid [PS] / https://idrid.grand-challenge.org/ [RESEARCHED]
   • Source Type: Official Academic Dataset & IEEE DataPort Repository [RESEARCHED].
   • Supported Claim: Provides 81 pixel-level annotated fundus images for MAs, hemorrhages, hard exudates, soft exudates, and optic disc, plus 516 graded images from an Indian population [RESEARCHED].
   • Reliability: High (Official SIH dataset source) [PS].

2. APTOS 2019 Blindness Detection Dataset

   • URL: https://www.kaggle.com/c/aptos2019-blindness-detection [PS]
   • Source Type: Kaggle Competition Dataset (Aravind Eye Hospital, India) [RESEARCHED].
   • Supported Claim: Provides 3,662 high-resolution retinal fundus images categorized into 5 ICDR severity levels from rural Indian screening camps [RESEARCHED].
   • Reliability: High (Official SIH dataset source) [PS].

3. Messidor-2 Dataset

   • URL: https://www.adcis.net/en/third-party/messidor2/ [PS]
   • Source Type: ADCIS / Brest University Hospital Research Repository [RESEARCHED].
   • Supported Claim: Provides 1,748 eye fundus images for evaluating diabetic retinopathy screening algorithms [RESEARCHED].
   • Reliability: High (Official SIH dataset source) [PS].

4. DRIVE: Digital Retinal Images for Vessel Extraction

   • URL: https://drive.grand-challenge.org/ [PS]
   • Source Type: Biomedical Imaging Grand Challenge Benchmark [RESEARCHED].
   • Supported Claim: Standard benchmark dataset containing 40 fundus images with manual blood vessel segmentations [RESEARCHED].
   • Reliability: High (Official SIH dataset source) [PS].

5. MathWorks Medical Imaging & Deep Learning Documentation

   • URL: https://www.mathworks.com/help/medical-imaging/ug/multilabel-diabetic-retinopathy-fundus-image-classification-using-deep-learning.html [RESEARCHED]
   • Source Type: Official MATLAB Documentation [RESEARCHED].
   • Supported Claim: Documents native MATLAB functions for fundus image classification, U-Net semantic segmentation, and gradCAM() explainability [RESEARCHED].
   • Reliability: High (Official vendor documentation) [RESEARCHED].